2 * Copyright (c) 2000, 2003 Silicon Graphics, Inc. All rights reserved.
3 * Copyright (c) 2001 Intel Corp.
4 * Copyright (c) 2001 Tony Luck <tony.luck@intel.com>
5 * Copyright (c) 2002 NEC Corp.
6 * Copyright (c) 2002 Kimio Suganuma <k-suganuma@da.jp.nec.com>
7 * Copyright (c) 2004 Silicon Graphics, Inc
8 * Russ Anderson <rja@sgi.com>
9 * Jesse Barnes <jbarnes@sgi.com>
10 * Jack Steiner <steiner@sgi.com>
14 * Platform initialization for Discontig Memory
17 #include <linux/kernel.h>
19 #include <linux/swap.h>
20 #include <linux/bootmem.h>
21 #include <linux/acpi.h>
22 #include <linux/efi.h>
23 #include <linux/nodemask.h>
24 #include <asm/pgalloc.h>
26 #include <asm/meminit.h>
28 #include <asm/sections.h>
31 * Track per-node information needed to setup the boot memory allocator, the
32 * per-node areas, and the real VM.
34 struct early_node_data {
35 struct ia64_node_data *node_data;
37 unsigned long pernode_addr;
38 unsigned long pernode_size;
39 struct bootmem_data bootmem_data;
40 unsigned long num_physpages;
41 unsigned long num_dma_physpages;
42 unsigned long min_pfn;
43 unsigned long max_pfn;
46 static struct early_node_data mem_data[MAX_NUMNODES] __initdata;
49 * reassign_cpu_only_nodes - called from find_memory to move CPU-only nodes to a memory node
51 * This function will move nodes with only CPUs (no memory)
52 * to a node with memory which is at the minimum numa_slit distance.
53 * Any reassigments will result in the compression of the nodes
54 * and renumbering the nid values where appropriate.
55 * The static declarations below are to avoid large stack size which
56 * makes the code not re-entrant.
58 static void __init reassign_cpu_only_nodes(void)
60 struct node_memblk_s *p;
61 int i, j, k, nnode, nid, cpu, cpunid, pxm;
63 static DECLARE_BITMAP(nodes_with_mem, MAX_NUMNODES) __initdata;
64 static u8 numa_slit_fix[MAX_NUMNODES * MAX_NUMNODES] __initdata;
65 static int node_flip[MAX_NUMNODES] __initdata;
66 static int old_nid_map[NR_CPUS] __initdata;
68 for (nnode = 0, p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
69 if (!test_bit(p->nid, (void *) nodes_with_mem)) {
70 set_bit(p->nid, (void *) nodes_with_mem);
75 * All nids with memory.
77 if (nnode == num_online_nodes())
81 * Change nids and attempt to migrate CPU-only nodes
82 * to the best numa_slit (closest neighbor) possible.
83 * For reassigned CPU nodes a nid can't be arrived at
84 * until after this loop because the target nid's new
85 * identity might not have been established yet. So
86 * new nid values are fabricated above num_online_nodes() and
87 * mapped back later to their true value.
89 /* MCD - This code is a bit complicated, but may be unnecessary now.
90 * We can now handle much more interesting node-numbering.
91 * The old requirement that 0 <= nid <= numnodes <= MAX_NUMNODES
92 * and that there be no holes in the numbering 0..numnodes
93 * has become simply 0 <= nid <= MAX_NUMNODES.
96 for_each_online_node(i) {
97 if (test_bit(i, (void *) nodes_with_mem)) {
99 * Save original nid value for numa_slit
100 * fixup and node_cpuid reassignments.
109 for (p = &node_memblk[0]; p < &node_memblk[num_node_memblks]; p++)
116 cpunid = MAX_NUMNODES;
118 for (cpu = 0; cpu < NR_CPUS; cpu++)
119 if (node_cpuid[cpu].nid == i) {
121 * For nodes not being reassigned just
122 * fix the cpu's nid and reverse pxm map
124 if (cpunid < MAX_NUMNODES) {
125 pxm = nid_to_pxm_map[i];
126 pxm_to_nid_map[pxm] =
127 node_cpuid[cpu].nid = cpunid;
132 * For nodes being reassigned, find best node by
133 * numa_slit information and then make a temporary
134 * nid value based on current nid and num_online_nodes().
137 k = 2*num_online_nodes();
138 for_each_online_node(j) {
141 else if (test_bit(j, (void *) nodes_with_mem)) {
142 cslit = numa_slit[i * num_online_nodes() + j];
144 k = num_online_nodes() + j;
150 /* save old nid map so we can update the pxm */
151 old_nid_map[cpu] = node_cpuid[cpu].nid;
152 node_cpuid[cpu].nid = k;
157 * Fixup temporary nid values for CPU-only nodes.
159 for (cpu = 0; cpu < NR_CPUS; cpu++)
160 if (node_cpuid[cpu].nid == (2*num_online_nodes())) {
161 pxm = nid_to_pxm_map[old_nid_map[cpu]];
162 pxm_to_nid_map[pxm] = node_cpuid[cpu].nid = nnode - 1;
164 for (i = 0; i < nnode; i++) {
165 if (node_flip[i] != (node_cpuid[cpu].nid - num_online_nodes()))
168 pxm = nid_to_pxm_map[old_nid_map[cpu]];
169 pxm_to_nid_map[pxm] = node_cpuid[cpu].nid = i;
175 * Fix numa_slit by compressing from larger
176 * nid array to reduced nid array.
178 for (i = 0; i < nnode; i++)
179 for (j = 0; j < nnode; j++)
180 numa_slit_fix[i * nnode + j] =
181 numa_slit[node_flip[i] * num_online_nodes() + node_flip[j]];
183 memcpy(numa_slit, numa_slit_fix, sizeof (numa_slit));
185 nodes_clear(node_online_map);
186 for (i = 0; i < nnode; i++)
193 * To prevent cache aliasing effects, align per-node structures so that they
194 * start at addresses that are strided by node number.
196 #define NODEDATA_ALIGN(addr, node) \
197 ((((addr) + 1024*1024-1) & ~(1024*1024-1)) + (node)*PERCPU_PAGE_SIZE)
200 * build_node_maps - callback to setup bootmem structs for each node
201 * @start: physical start of range
202 * @len: length of range
203 * @node: node where this range resides
205 * We allocate a struct bootmem_data for each piece of memory that we wish to
206 * treat as a virtually contiguous block (i.e. each node). Each such block
207 * must start on an %IA64_GRANULE_SIZE boundary, so we round the address down
208 * if necessary. Any non-existent pages will simply be part of the virtual
209 * memmap. We also update min_low_pfn and max_low_pfn here as we receive
210 * memory ranges from the caller.
212 static int __init build_node_maps(unsigned long start, unsigned long len,
215 unsigned long cstart, epfn, end = start + len;
216 struct bootmem_data *bdp = &mem_data[node].bootmem_data;
218 epfn = GRANULEROUNDUP(end) >> PAGE_SHIFT;
219 cstart = GRANULEROUNDDOWN(start);
221 if (!bdp->node_low_pfn) {
222 bdp->node_boot_start = cstart;
223 bdp->node_low_pfn = epfn;
225 bdp->node_boot_start = min(cstart, bdp->node_boot_start);
226 bdp->node_low_pfn = max(epfn, bdp->node_low_pfn);
229 min_low_pfn = min(min_low_pfn, bdp->node_boot_start>>PAGE_SHIFT);
230 max_low_pfn = max(max_low_pfn, bdp->node_low_pfn);
236 * early_nr_phys_cpus_node - return number of physical cpus on a given node
237 * @node: node to check
239 * Count the number of physical cpus on @node. These are cpus that actually
240 * exist. We can't use nr_cpus_node() yet because
241 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
244 static int early_nr_phys_cpus_node(int node)
248 for (cpu = 0; cpu < NR_CPUS; cpu++)
249 if (node == node_cpuid[cpu].nid)
250 if ((cpu == 0) || node_cpuid[cpu].phys_id)
258 * early_nr_cpus_node - return number of cpus on a given node
259 * @node: node to check
261 * Count the number of cpus on @node. We can't use nr_cpus_node() yet because
262 * acpi_boot_init() (which builds the node_to_cpu_mask array) hasn't been
263 * called yet. Note that node 0 will also count all non-existent cpus.
265 static int early_nr_cpus_node(int node)
269 for (cpu = 0; cpu < NR_CPUS; cpu++)
270 if (node == node_cpuid[cpu].nid)
277 * find_pernode_space - allocate memory for memory map and per-node structures
278 * @start: physical start of range
279 * @len: length of range
280 * @node: node where this range resides
282 * This routine reserves space for the per-cpu data struct, the list of
283 * pg_data_ts and the per-node data struct. Each node will have something like
284 * the following in the first chunk of addr. space large enough to hold it.
286 * ________________________
288 * |~~~~~~~~~~~~~~~~~~~~~~~~| <-- NODEDATA_ALIGN(start, node) for the first
289 * | PERCPU_PAGE_SIZE * | start and length big enough
290 * | cpus_on_this_node | Node 0 will also have entries for all non-existent cpus.
291 * |------------------------|
292 * | local pg_data_t * |
293 * |------------------------|
294 * | local ia64_node_data |
295 * |------------------------|
297 * |________________________|
299 * Once this space has been set aside, the bootmem maps are initialized. We
300 * could probably move the allocation of the per-cpu and ia64_node_data space
301 * outside of this function and use alloc_bootmem_node(), but doing it here
302 * is straightforward and we get the alignments we want so...
304 static int __init find_pernode_space(unsigned long start, unsigned long len,
307 unsigned long epfn, cpu, cpus, phys_cpus;
308 unsigned long pernodesize = 0, pernode, pages, mapsize;
310 struct bootmem_data *bdp = &mem_data[node].bootmem_data;
312 epfn = (start + len) >> PAGE_SHIFT;
314 pages = bdp->node_low_pfn - (bdp->node_boot_start >> PAGE_SHIFT);
315 mapsize = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
318 * Make sure this memory falls within this node's usable memory
319 * since we may have thrown some away in build_maps().
321 if (start < bdp->node_boot_start || epfn > bdp->node_low_pfn)
324 /* Don't setup this node's local space twice... */
325 if (mem_data[node].pernode_addr)
329 * Calculate total size needed, incl. what's necessary
330 * for good alignment and alias prevention.
332 cpus = early_nr_cpus_node(node);
333 phys_cpus = early_nr_phys_cpus_node(node);
334 pernodesize += PERCPU_PAGE_SIZE * cpus;
335 pernodesize += node * L1_CACHE_BYTES;
336 pernodesize += L1_CACHE_ALIGN(sizeof(pg_data_t));
337 pernodesize += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
338 pernodesize = PAGE_ALIGN(pernodesize);
339 pernode = NODEDATA_ALIGN(start, node);
341 /* Is this range big enough for what we want to store here? */
342 if (start + len > (pernode + pernodesize + mapsize)) {
343 mem_data[node].pernode_addr = pernode;
344 mem_data[node].pernode_size = pernodesize;
345 memset(__va(pernode), 0, pernodesize);
347 cpu_data = (void *)pernode;
348 pernode += PERCPU_PAGE_SIZE * cpus;
349 pernode += node * L1_CACHE_BYTES;
351 mem_data[node].pgdat = __va(pernode);
352 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
354 mem_data[node].node_data = __va(pernode);
355 pernode += L1_CACHE_ALIGN(sizeof(struct ia64_node_data));
357 mem_data[node].pgdat->bdata = bdp;
358 pernode += L1_CACHE_ALIGN(sizeof(pg_data_t));
361 * Copy the static per-cpu data into the region we
362 * just set aside and then setup __per_cpu_offset
363 * for each CPU on this node.
365 for (cpu = 0; cpu < NR_CPUS; cpu++) {
366 if (node == node_cpuid[cpu].nid) {
367 memcpy(__va(cpu_data), __phys_per_cpu_start,
368 __per_cpu_end - __per_cpu_start);
369 __per_cpu_offset[cpu] = (char*)__va(cpu_data) -
371 cpu_data += PERCPU_PAGE_SIZE;
380 * free_node_bootmem - free bootmem allocator memory for use
381 * @start: physical start of range
382 * @len: length of range
383 * @node: node where this range resides
385 * Simply calls the bootmem allocator to free the specified ranged from
386 * the given pg_data_t's bdata struct. After this function has been called
387 * for all the entries in the EFI memory map, the bootmem allocator will
388 * be ready to service allocation requests.
390 static int __init free_node_bootmem(unsigned long start, unsigned long len,
393 free_bootmem_node(mem_data[node].pgdat, start, len);
399 * reserve_pernode_space - reserve memory for per-node space
401 * Reserve the space used by the bootmem maps & per-node space in the boot
402 * allocator so that when we actually create the real mem maps we don't
405 static void __init reserve_pernode_space(void)
407 unsigned long base, size, pages;
408 struct bootmem_data *bdp;
411 for_each_online_node(node) {
412 pg_data_t *pdp = mem_data[node].pgdat;
416 /* First the bootmem_map itself */
417 pages = bdp->node_low_pfn - (bdp->node_boot_start>>PAGE_SHIFT);
418 size = bootmem_bootmap_pages(pages) << PAGE_SHIFT;
419 base = __pa(bdp->node_bootmem_map);
420 reserve_bootmem_node(pdp, base, size);
422 /* Now the per-node space */
423 size = mem_data[node].pernode_size;
424 base = __pa(mem_data[node].pernode_addr);
425 reserve_bootmem_node(pdp, base, size);
430 * initialize_pernode_data - fixup per-cpu & per-node pointers
432 * Each node's per-node area has a copy of the global pg_data_t list, so
433 * we copy that to each node here, as well as setting the per-cpu pointer
434 * to the local node data structure. The active_cpus field of the per-node
435 * structure gets setup by the platform_cpu_init() function later.
437 static void __init initialize_pernode_data(void)
440 pg_data_t *pgdat_list[MAX_NUMNODES];
442 for_each_online_node(node)
443 pgdat_list[node] = mem_data[node].pgdat;
445 /* Copy the pg_data_t list to each node and init the node field */
446 for_each_online_node(node) {
447 memcpy(mem_data[node].node_data->pg_data_ptrs, pgdat_list,
451 /* Set the node_data pointer for each per-cpu struct */
452 for (cpu = 0; cpu < NR_CPUS; cpu++) {
453 node = node_cpuid[cpu].nid;
454 per_cpu(cpu_info, cpu).node_data = mem_data[node].node_data;
459 * find_memory - walk the EFI memory map and setup the bootmem allocator
461 * Called early in boot to setup the bootmem allocator, and to
462 * allocate the per-cpu and per-node structures.
464 void __init find_memory(void)
470 if (num_online_nodes() == 0) {
471 printk(KERN_ERR "node info missing!\n");
478 if (num_online_nodes() > 1)
479 reassign_cpu_only_nodes();
481 /* These actually end up getting called by call_pernode_memory() */
482 efi_memmap_walk(filter_rsvd_memory, build_node_maps);
483 efi_memmap_walk(filter_rsvd_memory, find_pernode_space);
486 * Initialize the boot memory maps in reverse order since that's
487 * what the bootmem allocator expects
489 for (node = MAX_NUMNODES - 1; node >= 0; node--) {
490 unsigned long pernode, pernodesize, map;
491 struct bootmem_data *bdp;
493 if (!node_online(node))
496 bdp = &mem_data[node].bootmem_data;
497 pernode = mem_data[node].pernode_addr;
498 pernodesize = mem_data[node].pernode_size;
499 map = pernode + pernodesize;
501 /* Sanity check... */
503 panic("pernode space for node %d "
504 "could not be allocated!", node);
506 init_bootmem_node(mem_data[node].pgdat,
508 bdp->node_boot_start>>PAGE_SHIFT,
512 efi_memmap_walk(filter_rsvd_memory, free_node_bootmem);
514 reserve_pernode_space();
515 initialize_pernode_data();
517 max_pfn = max_low_pfn;
523 * per_cpu_init - setup per-cpu variables
525 * find_pernode_space() does most of this already, we just need to set
526 * local_per_cpu_offset
528 void *per_cpu_init(void)
532 if (smp_processor_id() == 0) {
533 for (cpu = 0; cpu < NR_CPUS; cpu++) {
534 per_cpu(local_per_cpu_offset, cpu) =
535 __per_cpu_offset[cpu];
539 return __per_cpu_start + __per_cpu_offset[smp_processor_id()];
543 * show_mem - give short summary of memory stats
545 * Shows a simple page count of reserved and used pages in the system.
546 * For discontig machines, it does this on a per-pgdat basis.
550 int i, total_reserved = 0;
551 int total_shared = 0, total_cached = 0;
552 unsigned long total_present = 0;
555 printk("Mem-info:\n");
557 printk("Free swap: %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
558 for_each_pgdat(pgdat) {
559 unsigned long present = pgdat->node_present_pages;
560 int shared = 0, cached = 0, reserved = 0;
561 printk("Node ID: %d\n", pgdat->node_id);
562 for(i = 0; i < pgdat->node_spanned_pages; i++) {
563 struct page *page = pgdat_page_nr(pgdat, i);
564 if (!ia64_pfn_valid(pgdat->node_start_pfn+i))
566 if (PageReserved(page))
568 else if (PageSwapCache(page))
570 else if (page_count(page))
571 shared += page_count(page)-1;
573 total_present += present;
574 total_reserved += reserved;
575 total_cached += cached;
576 total_shared += shared;
577 printk("\t%ld pages of RAM\n", present);
578 printk("\t%d reserved pages\n", reserved);
579 printk("\t%d pages shared\n", shared);
580 printk("\t%d pages swap cached\n", cached);
582 printk("%ld pages of RAM\n", total_present);
583 printk("%d reserved pages\n", total_reserved);
584 printk("%d pages shared\n", total_shared);
585 printk("%d pages swap cached\n", total_cached);
586 printk("Total of %ld pages in page table cache\n",
587 pgtable_quicklist_total_size());
588 printk("%d free buffer pages\n", nr_free_buffer_pages());
592 * call_pernode_memory - use SRAT to call callback functions with node info
593 * @start: physical start of range
594 * @len: length of range
595 * @arg: function to call for each range
597 * efi_memmap_walk() knows nothing about layout of memory across nodes. Find
598 * out to which node a block of memory belongs. Ignore memory that we cannot
599 * identify, and split blocks that run across multiple nodes.
601 * Take this opportunity to round the start address up and the end address
602 * down to page boundaries.
604 void call_pernode_memory(unsigned long start, unsigned long len, void *arg)
606 unsigned long rs, re, end = start + len;
607 void (*func)(unsigned long, unsigned long, int);
610 start = PAGE_ALIGN(start);
617 if (!num_node_memblks) {
618 /* No SRAT table, so assume one node (node 0) */
620 (*func)(start, end - start, 0);
624 for (i = 0; i < num_node_memblks; i++) {
625 rs = max(start, node_memblk[i].start_paddr);
626 re = min(end, node_memblk[i].start_paddr +
627 node_memblk[i].size);
630 (*func)(rs, re - rs, node_memblk[i].nid);
638 * count_node_pages - callback to build per-node memory info structures
639 * @start: physical start of range
640 * @len: length of range
641 * @node: node where this range resides
643 * Each node has it's own number of physical pages, DMAable pages, start, and
644 * end page frame number. This routine will be called by call_pernode_memory()
645 * for each piece of usable memory and will setup these values for each node.
646 * Very similar to build_maps().
648 static __init int count_node_pages(unsigned long start, unsigned long len, int node)
650 unsigned long end = start + len;
652 mem_data[node].num_physpages += len >> PAGE_SHIFT;
653 if (start <= __pa(MAX_DMA_ADDRESS))
654 mem_data[node].num_dma_physpages +=
655 (min(end, __pa(MAX_DMA_ADDRESS)) - start) >>PAGE_SHIFT;
656 start = GRANULEROUNDDOWN(start);
657 start = ORDERROUNDDOWN(start);
658 end = GRANULEROUNDUP(end);
659 mem_data[node].max_pfn = max(mem_data[node].max_pfn,
661 mem_data[node].min_pfn = min(mem_data[node].min_pfn,
662 start >> PAGE_SHIFT);
668 * paging_init - setup page tables
670 * paging_init() sets up the page tables for each node of the system and frees
671 * the bootmem allocator memory for general use.
673 void __init paging_init(void)
675 unsigned long max_dma;
676 unsigned long zones_size[MAX_NR_ZONES];
677 unsigned long zholes_size[MAX_NR_ZONES];
678 unsigned long pfn_offset = 0;
681 max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
683 /* so min() will work in count_node_pages */
684 for_each_online_node(node)
685 mem_data[node].min_pfn = ~0UL;
687 efi_memmap_walk(filter_rsvd_memory, count_node_pages);
689 for_each_online_node(node) {
690 memset(zones_size, 0, sizeof(zones_size));
691 memset(zholes_size, 0, sizeof(zholes_size));
693 num_physpages += mem_data[node].num_physpages;
695 if (mem_data[node].min_pfn >= max_dma) {
696 /* All of this node's memory is above ZONE_DMA */
697 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
698 mem_data[node].min_pfn;
699 zholes_size[ZONE_NORMAL] = mem_data[node].max_pfn -
700 mem_data[node].min_pfn -
701 mem_data[node].num_physpages;
702 } else if (mem_data[node].max_pfn < max_dma) {
703 /* All of this node's memory is in ZONE_DMA */
704 zones_size[ZONE_DMA] = mem_data[node].max_pfn -
705 mem_data[node].min_pfn;
706 zholes_size[ZONE_DMA] = mem_data[node].max_pfn -
707 mem_data[node].min_pfn -
708 mem_data[node].num_dma_physpages;
710 /* This node has memory in both zones */
711 zones_size[ZONE_DMA] = max_dma -
712 mem_data[node].min_pfn;
713 zholes_size[ZONE_DMA] = zones_size[ZONE_DMA] -
714 mem_data[node].num_dma_physpages;
715 zones_size[ZONE_NORMAL] = mem_data[node].max_pfn -
717 zholes_size[ZONE_NORMAL] = zones_size[ZONE_NORMAL] -
718 (mem_data[node].num_physpages -
719 mem_data[node].num_dma_physpages);
724 PAGE_ALIGN(max_low_pfn * sizeof(struct page));
725 vmem_map = (struct page *) vmalloc_end;
727 efi_memmap_walk(create_mem_map_page_table, NULL);
728 printk("Virtual mem_map starts at 0x%p\n", vmem_map);
731 pfn_offset = mem_data[node].min_pfn;
733 NODE_DATA(node)->node_mem_map = vmem_map + pfn_offset;
734 free_area_init_node(node, NODE_DATA(node), zones_size,
735 pfn_offset, zholes_size);
738 zero_page_memmap_ptr = virt_to_page(ia64_imva(empty_zero_page));